The network below runs successfully on CPU, but throws a RuntimeError: Can only downcast contiguous tensors when backpropagating through the AvgPool3d layer on CUDA. If either the initial or final Conv3d layers are removed, the network runs successfully. Any explanations of why this is happening or suggestions on how to fix it would be appreciated. I am using pytorch 0.2.0 on python 3.5 with CUDA 8.0 from the pip binary, on ubuntu 16.04.
import torch
from torch import nn
from torch.autograd import Variable
class ConvBlock(nn.Module):
def __init__(self, in_channels, out_channels):
super(ConvBlock, self).__init__()
self.net = nn.Conv3d(in_channels=in_channels, out_channels=out_channels,
kernel_size=(3, 5, 5), stride=(3, 3, 3),
padding=(1, 0, 0), bias=False)
def forward(self, x):
return self.net(x)
class DepthConcatBlock(nn.Module):
def __init__(self, channels):
super(DepthConcatBlock, self).__init__()
self.branch1 = nn.Conv3d(in_channels=channels, out_channels=channels,
kernel_size=(1, 3, 3), stride=(1, 2, 2), padding=0)
self.branch2 = nn.AvgPool3d(kernel_size=(1, 3, 3), stride=(1, 2, 2))
def forward(self, x):
return torch.cat([self.branch1(x), self.branch2(x)], dim=1)
net = nn.Sequential(
ConvBlock(3, 3),
DepthConcatBlock(3),
ConvBlock(6, 6)
)
net = net.cuda()
x = Variable(torch.randn([1, 3, 75, 288, 360])).cuda()
y = net(x)
l = torch.sum(y)
l.backward()
Traceback (most recent call last):
File "debug.py", line 39, in <module>
l.backward()
File "/usr/local/lib/python3.5/dist-packages/torch/autograd/variable.py", line 156, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, retain_variables)
File "/usr/local/lib/python3.5/dist-packages/torch/autograd/__init__.py", line 98, in backward
variables, grad_variables, retain_graph)
File "/usr/local/lib/python3.5/dist-packages/torch/autograd/function.py", line 91, in apply
return self._forward_cls.backward(self, *args)
File "/usr/local/lib/python3.5/dist-packages/torch/nn/_functions/thnn/pooling.py", line 419, in backward
grad_input = AvgPool3dBackward.apply(input, grad_output, ctx.kernel_size, ctx.stride)
File "/usr/local/lib/python3.5/dist-packages/torch/nn/_functions/thnn/pooling.py", line 435, in forward
ctx.stride[0], ctx.stride[2], ctx.stride[1])
RuntimeError: Can only downcast contiguous tensors at /pytorch/torch/lib/tmp_install/include/THC/THCDeviceTensor-inl.cuh:295